Audit-Grade Prompt Lineage: An Analytical Study on End-to-End Traceability for Snowflake Cortex LLM Functions in Regulated Reporting

Audit-Grade Prompt Lineage

Abstract

Large Language Models have moved quickly from proof-of-concept tools into the operational core of enterprise reporting. The harder question in regulated industries is whether organizations can demonstrate, with the evidentiary rigour that auditors require, how a specific statement was produced, what data shaped it, and whether the process could be independently verified.

This study examines the traceability gap that emerges when Snowflake Cortex LLM functions are embedded inside regulated reporting pipelines and proposes a formal end-to-end prompt lineage framework. The framework is evaluated through scenario-based audit simulation against five regulatory examination scenarios and requirements mapping across six major compliance frameworks: EU AI Act, SOX, SEC AI disclosure guidance, Basel IV, DORA, and NIST AI RMF.

A gap analysis confirms that prompt construction capture, output hashing, and report-section mapping show near-zero coverage in current enterprise governance tooling. Scenario-based validation demonstrates that the proposed framework produces examiner-ready responses across all five audit scenarios where existing tools consistently fail.

The paper establishes prompt lineage as a foundational governance requirement for regulated industries deploying LLMs in production reporting pipelines, offering a deployable reference architecture that organizations can adopt before their next regulatory examination.